AWS just wrapped its Sydney summit. The headline number was nine new renewable power purchase agreements totalling 430MW, taking Amazon’s Australian renewable capacity to nearly 1GW. Behind it sits a AU$20 billion commitment to data centre infrastructure by 2029.

That reads as ambition. It is also partial cover for a problem the keynote slides do not show: the grid is full, and Australia’s biggest hyperscaler cannot buy its way out of that fast enough.

The number that actually matters

In the last 18 months, NSW transmission operator Transgrid received more than 10GW of potential data centre load enquiries. Individual projects range from 250MW to over 1GW. For context, that 10GW pipeline is more than half of the entire state’s peak electricity demand.

NSW and Victoria already host around 80% of Australia’s data centre capacity. The compute is concentrated in two metros. The grid in those two metros was not built for it.

AirTrunk’s proposed SYD4 facility in Western Sydney is 1.2GW, the first Australian data centre to clear the gigawatt threshold. NextDC has a $7 billion 550MW build in the same corridor with OpenAI as anchor tenant. Each new announcement competes for the same substations, the same transmission corridors, the same firmed renewable supply.

The PPA pipeline matters. The substation queue matters more.

What “stalled” actually looks like

Grid connection in Australia is not a procurement step. It is a multi-year capital project. Connections to essential utilities (power, water, sewerage) are now the most prominent delays to project delivery, running anywhere from one to five years depending on interconnection queues and underlying generation.

The queue is also lying to itself. Oxford Economics estimates that six in every seven megawatts in connection requests is “phantom demand” that will not materialise. Of the 44GW of requests AEMO received in its 2025 scenario report, only about 6GW reflects actual required capacity. Developers reserve at multiple utilities to lock in optionality. That makes prioritisation harder and stretches lead times further.

On 12 March 2026, the AEMC proposed lifting the threshold for large inverter-based load connections from 5MW to 30MW and tightening ride-through standards. The reason is operational, not bureaucratic. A single hyperscale facility tripping during a voltage event now moves the grid.

Then on 12 May 2026, federal and state energy ministers agreed data centres must fully offset their demand through new renewable generation and firming capacity, provide flexibility services, and report transparently on energy use. Queensland did not sign on. Transgrid’s position in the same announcement was direct: “Existing electricity consumers must be no worse off… this principle is non-negotiable.”

That is regulator language for: you are not putting this on household bills.

Why the AWS PPAs don’t solve it

Power purchase agreements buy generation. They do not build transmission, they do not energise substations, and they do not move a connection queue. Eight of Amazon’s nine new deals include battery storage, which helps firm intermittent renewables. None of it shortens the wait for a 100kV connection in Western Sydney.

This is the part the keynote does not explain. An AU$20 billion infrastructure commitment presumes the megawatts can be delivered to the racks. For the largest builds, that is not settled. The Investment Delivery Authority endorsed 15 data centre projects worth A$51.9 billion in March for prioritised approvals: a fast-track built for exactly this bottleneck. The fact a fast-track exists tells you what the slow track looks like.

What this means if you’re building

If you are a founder running an AI-native workload in Australia, three things change.

First, power availability is now an architectural decision, not a procurement footnote. “We’ll move to AWS Sydney when we scale” assumes Sydney has the headroom when you arrive. For training-heavy or high-density inference workloads, that is no longer safe to assume in 2027 and 2028. Multi-region from day one, with at least one region outside NSW and Victoria, is the defensive position.

Second, hyperscaler co-location is not a guaranteed cost line. The federal government’s stated expectation is that operators pay their full share of new grid connectivity so costs are not passed to consumers or businesses. That cost flows somewhere: into hyperscaler rate cards, into customer compute prices, or into capped facilities that raise the effective price per GPU-hour. None of it goes away.

Third, the model you train on may not be served from where you think. AWS, like every hyperscaler, will route inference where capacity exists. If Sydney runs hot, your “Australian” workload may quietly hop to Singapore or Jakarta — fine for latency-tolerant batch jobs, a sovereignty problem for regulated data. Read your residency contracts before you assume the cloud region badge means what it used to mean.

The contrarian read

The dominant narrative in Australian AI infrastructure right now is demand-side. There is not enough compute, not enough models, not enough talent. The Sydney keynote leaned into all of it. The harder constraint is upstream and physical. You can buy more GPUs. You cannot buy a faster substation upgrade.

AWS’s nine PPAs are real. The 1GW portfolio is real. The AU$20 billion commitment is real. All of it lands inside a transmission and connection system that takes years to expand, in two metros doing 80% of the work, with a regulator that has just told the industry to stop pushing costs onto households.

Founders treating Australian cloud capacity as elastic through 2028 are reading a market that no longer exists. The companies that win the AI infrastructure race here will be the ones who designed for the power constraint, not the ones who assumed it would solve itself.

If you are working out how to build AI-native infrastructure in a power-constrained market, talk to us.

Sources